A Simple Framework for Cross-Domain Few-Shot Recognition with Unlabeled Data

Most existing works in few-shot learning rely on metalearning the model on a large base dataset which is typically from the same domain as the target dataset. We tackle the problem of cross-domain few-shot learning where there is a large shift between the base and target domain. We propose a simple solution to utilize unlabeled images from the novel/base dataset. We calculate pseudo soft-label from the weakly-augmented version of the unlabeled image and compare it with the strongly augmented version. We also minimize the supervised cross-entropy loss for the labeled base dataset at the same time. We show that the proposed network learns representation that can be easily adapted to the target domain even though it has not been trained with target-specific classes during the pretraining phase. Our model outperforms the current state-of-the art method by 2.7% for 5-shot and 3.6% for 1-shot classification in the BSCD-FSL benchmark.